In recent years, online car is an important part of smart cities. Accurate demand forecasts allow drivers and passengers to better determine their travel itinerary, departure time and travel origin options, which can help with traffic management. Recently, a convolution-gated recurrence unit (ConvGRU) method was developed which is designed to model such motion based on deep learning techniques. Despite its good performance, ConvGRU may produce blurred extrapolated images and cannot achieve multimodal and skewed intensity distributions. To overcome these limitations, a generative confrontational ConvGRU (GA-ConvGRU) model was used in this paper. The model consists of two antagonistic learning systems, a ConvGRU-based generator and a convolutional neural network-based discriminator. Train these two systems with minimax games. Through the confrontational learning program, GA-ConvGRU can produce more realistic and more accurate extrapolation. This paper uses this method to predict the hot spot distribution area of short-term network car service demand. The results show that the average error rate of this method is in a very low range.

CICTP